H-Z-Ning/Senti-RoBERTa-Mini

Chinese Movie Review Sentiment Classification Model (5-Star Rating)

---

1. Model Overview

H-Z-Ning/Senti-RoBERTa-Mini is a lightweight Chinese RoBERTa model fine-tuned specifically for assigning 1-to-5-star sentiment ratings to Chinese movie short reviews. Built on the HFL-Tencent hfl/chinese-roberta-wwm-ext checkpoint, it retains a small footprint and fast inference, making it ideal for resource-constrained deployments.

---

2. Model Facts

Item	Details
Task	Chinese text classification (sentiment / star rating)
Labels	5 classes (1 star – 5 stars)
Base model	hfl/chinese-roberta-wwm-ext
Dataset	Kaggle: Douban Movie Short Comments (2000 K)
Training framework	🤗 transformers + Trainer
Language	Simplified Chinese
Parameters	≈ 102 M (same as base model)

---

3. Quick Start

3.1 Install Dependencies

pip install transformers torch

3.2 One-Line Inference

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

repo = "H-Z-Ning/Senti-RoBERTa-Mini"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForSequenceClassification.from_pretrained(repo)

text = "这个导演真厉害。"
inputs = tok(text, return_tensors="pt", truncation=True, max_length=256)
with torch.no_grad():
    logits = model(**inputs).logits
pred = int(torch.argmax(logits, dim=-1).item()) + 1  # 1..5
print("predicted rating:", pred)

---

4.Training source code

senti-roberta-mini training source code

5. Training Details

Hyper-parameter	Value
Base model	hfl/chinese-roberta-wwm-ext
Training framework	🤗 transformers Trainer
Training set	150 000 samples (randomly drawn from 2000 K)
Validation set	15 000 samples (same random draw)
Test set	full original test set
Max sequence length	256
Training epochs	3
Batch size	32 (train) / 64 (eval)
Learning rate	2e-5
Optimizer	AdamW
Weight decay	0.01
Scheduler	linear warmup (warmup_ratio=0.1)
Precision	FP16
Best-model criterion	QWK (↑)
Training time	≈ 120 min on single P100 (FP16)
Logging interval	every 10 steps

---

6. Citation

@misc{senti-roberta-mini-2025,
  title={Senti-RoBERTa-Mini: A Mini Chinese RoBERTa for Movie Review Rating},
  author={H-Z-Ning},
  year={2025},
  howpublished={\url{https://huggingface.co/H-Z-Ning/Senti-RoBERTa-Mini}}
}

---

7. License

This model is released under Apache-2.0. The base model hfl/chinese-roberta-wwm-ext is also Apache-2.0.

---

Community contributions and feedback are welcome! If you encounter any issues, please open an Issue or email the author.